On Wed, 2014-10-22 at 11:35 +0200, Hannes Frederic Sowa wrote:

On Mi, 2014-10-22 at 00:44 +0100, Ben Hutchings wrote:

quoted

There are several ways that VMs can take advantage of UFO and get the
host to do fragmentation for them:

drivers/net/macvtap.c:                  gso_type = SKB_GSO_UDP;
drivers/net/tun.c:                      skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
drivers/net/virtio_net.c:                       skb_shinfo(skb)->gso_type = SKB_GSO_UDP;

Our implementation of UFO for IPv6 does:

		fptr = (struct frag_hdr *)(skb_network_header(skb) + unfrag_ip6hlen);
		fptr->nexthdr = nexthdr;
		fptr->reserved = 0;
		fptr->identification = skb_shinfo(skb)->ip6_frag_id;

which assumes ip6_frag_id has been set.  That's only true if the local
stack constructed the skb; otherwise it appears we get zero.

This seems to be a regression as a result of:

commit 916e4cf46d0204806c062c8c6c4d1f633852c5b6
Author: Hannes Frederic Sowa [off-list ref]
Date:   Fri Feb 21 02:55:35 2014 +0100

    ipv6: reuse ip6_frag_id from ip6_ufo_append_data

However, that change seems reasonable - we *shouldn't* be choosing IDs
for any other stack.  Any paravirt net driver that can use IPv6 UFO
needs to have some way of passing a fragmentation ID to put in
skb_shared_info::ip6_frag_id.

Do we really gain a lot of performance by enabling UFO on those devices
or would it make sense to just drop support? It only helps fragmenting
large UDP packets, so I don't think it is worth it.

It's not been important enough for anyone to bother implementing it in
hardware/firmware aside from Neterion.

I'll shortly post patches to disable it.

Ben.

Otherwise I agree with Ben, we need to pass a fragmentation id from the
host over to the system segmenting the gso frame. Fragmentation ids must
be generated by the end system.

Hmm...

-- 
Ben Hutchings
Theory and practice are closer in theory than in practice.
                                - John Levine, moderator of comp.compilers